Pith. sign in

REVIEW 3 major objections 5 minor 75 references

Enhancing Fatigue Detection through Heterogeneous Multi-Source Data Integration and Cross-Domain Modality Imputation

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read A proof-of-concept framework shows that field fatigue detection can be improved by imputing lab-only EEG and ECG from shared wearable signals, raising accuracy from 82.3% to 92.9% on the VPFD dataset.

desk verdict A genuinely new problem setting and a plausible proof-of-concept, but the theoretical bound is wrong and the evaluation split does not yet rule out subject-specific confounds. read the letter →

arxiv 2507.16859 v5 pith:4QR5WXNY submitted 2025-07-21 cs.RO cs.AI

classification cs.ROcs.AI
keywords fatiguedetectioncross-domainmodalityimputationmulti-sourcedomainadaptationwearablephysiologicalsignalsEEGandECGheterogeneoussensorspacesgeneralizationbound
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper is trying to establish that a fatigue-detection system deployed in the field with only wearable sensors (heart rate, photoplethysmography, galvanic skin response, skin temperature, acceleration) can be made more accurate by imputing the high-fidelity signals it lacks—electroencephalography and electrocardiography—using models trained on laboratory datasets that recorded those signals alongside the same wearables. The authors formalize this 'sensor-constrained' setting and propose a heterogeneous multi-source framework: shared wearable modalities form a bridge, source-domain imputers map them to the missing EEG/ECG, and the imputed signals are appended to the target features before classification. As a proof of concept with deliberately standard components, the best configuration (TimeMixer imputer, Transformer classifier) raises target-domain accuracy from 82.31% to 92.86% on the VPFD dataset, supporting the claim that cross-domain modality imputation can carry fatigue-relevant information from controlled lab environments to practical deployments. The authors position this as distinct from same-domain imputation and from standard cross-domain adaptation, since an entire modality is never observed in the target domain.

What carries the argument

The load-bearing object is the cross-domain modality imputer: a regression model $g_s: \mathcal{X}^\cap \to \mathcal{X}^{s+}$ trained on source-domain data to predict a source-only, hard-to-deploy modality from the modalities it shares with the target domain. Its job is to produce an estimate $\hat{a}_T$ of the missing EEG/ECG from the target's own wearable signals, so the target detector can use the augmented feature vector $[x^\cap, \hat{a}]$. The imputer's usefulness is justified by three theorems: adding an informative modality tightens a mutual-information generalization bound (Theorem 1), target error with imputed data is controlled by the $\mathcal{H}\Delta\mathcal{H}$ distance between real and imputed modalities (Theorem 2), and that distance is in turn bounded by the source imputation error plus the domain gap in shared features (Theorem 3). Two regularizers implement the alignment suggested by Theorem 3: batch normalization is applied to the imputer and classifier, and Jacobian-norm regularization is applied to classifiers. In the reported best configuration the imputer is TimeMixer and the classifier is a Transformer, with multi-source information combined by simple sequential concatenation.

What would settle it

Equip a small field-like cohort with both the wearable set and a reference EEG/ECG recorder, then compare the imputer's outputs with the actual recordings (e.g., correlation coefficient or $\mathcal{H}\Delta\mathcal{H}$ distance) and rerun the classification using real versus imputed modalities; if the accuracy gain disappears when imputed signals are replaced by noise-matched random signals, the transfer is not carrying fatigue information.

Watch

Extended reading notes

Core claim

The paper claims that fatigue-relevant information can be transferred across differently instrumented domains through shared wearable modalities, and that this transfer measurably improves fatigue detection when high-fidelity sensors are absent from the deployment environment. Concretely, the framework trains a modality imputer $g_s$ on each source domain, using the modalities shared with the target (HR, PPG, GSR, ST, ACC) as inputs and the source-only high-fidelity modality (EEG from MEFAR, ECG from FatigueSet) as the prediction target; the trained imputer is then applied to the target's shared features, and the imputed signals are concatenated to the target feature vector before a classifier is trained. On the VPFD target dataset, the best reported configuration—sequential imputation of ECG then EEG with TimeMixer, classified by a Transformer—reaches 92.86% accuracy against an 82.31% baseline for the same classifier on unmodified VPFD, with the EEG+ECG combination generally outperforming single-modality imputation. The authors also prove generalization bounds (Theorems 1–3) showing that adding an imputed, label-related modality tightens the error bound provided the imputed distribution is close to the real one in $\mathcal{H}\Delta\mathcal{H}$ distance and the shared-feature distributions are aligned.

Load-bearing premise

The relationship between shared wearable signals and the missing EEG/ECG signals, learned on laboratory data from different subjects and devices, is assumed to carry over to the field target domain, and this assumption is never checked against real EEG/ECG recordings in the target setting.

Editorial extensions

If this is right

  • Field fatigue detectors can be upgraded with EEG/ECG-based knowledge without ever purchasing or operating EEG/ECG sensors, as long as they record at least the shared wearable modalities.
  • Public lab datasets that happen to record overlapping wearable channels can serve as reusable sensor-knowledge banks for many different target deployments.
  • Combining multiple source domains, each contributing a different missing modality, yields complementary gains over using any single source alone.
  • Imputer selection matters more than raw reconstruction error: the MLP imputer had the lowest MSE but the TimeMixer imputer produced the highest downstream accuracy, so imputers should be evaluated by end-task performance.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If the transfer mechanism is genuine, the same recipe could impute other deployment-constrained modalities—for example pupil dilation from wearable signals alone—using any dataset that recorded both.
  • The accuracy gains are only validated through downstream classification, not through physiological fidelity; the framework would be stronger if imputed EEG/ECG were compared with concurrently recorded reference signals in a target-like setting.
  • The reported gain might partly reflect classifier-specific interactions rather than information content, since the MSE-optimal imputer did not yield the best classification; a direct test would be to fix the classifier and vary only imputed-signal quality.
  • The multi-source fusion step is deliberately naive (sequential concatenation without conflict weighting); introducing adaptive weighting when sources disagree could push performance further, a direction the paper lists as future work.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper proposes a framework for fatigue detection in a sensor-constrained target domain by imputing missing high-fidelity modalities (EEG, ECG) from source domains (MEFAR, FatigueSet) through shared wearable modalities (HR, PPG, GSR, ST, ACC). The imputer is trained on source-domain pairs of shared and target modalities, then applied to target-domain shared features; the augmented features are used to train classifiers on VPFD. Experiments compare several imputers (MLP, LSTM, CNN1D, TimeMixer) and classifiers (MLP, LSTM, CNN1D, Transformer), reporting accuracy gains, e.g., 92.86% for the Transformer with TimeMixer-imputed EEG+ECG versus 82.31% on the VPFD baseline. A theoretical section (Theorems 1–3) is used to motivate the approach, claiming that adding label-related modalities tightens the generalization bound and that imputation quality depends on distributional distances.

Significance. The paper addresses a genuinely practical problem: transferring fatigue-relevant knowledge from laboratory-grade sensors to field-deployable wearable setups where high-fidelity modalities are absent. It formalizes the heterogeneous multi-source setting, uses external public datasets, and the imputer is trained without target labels, so the empirical evaluation is not circular. The use of multiple standard architectures as a proof-of-concept is a strength, as is the explicit discussion of limitations in the conclusion. If the reported gains survive subject-independent evaluation, the framework would be a useful contribution to cross-domain physiological-signal transfer. However, the current evidence is not sufficient to establish the central claim of cross-domain transfer because the evaluation protocol allows within-subject temporal correlation to explain the gains, and the theoretical motivation contains a sign error that invalidates Theorem 1 as stated.

major comments (3)
  1. [Appendix, Theorem 1 (Eqs. 7–13)] The proof of Theorem 1 asserts I(x+, y) - I(x, y) = H(Y|x) - H(Y|x+) < 0, but conditioning on more variables cannot increase conditional entropy: since x+ = [x, a] contains at least as much information as x, H(Y|x) >= H(Y|x+), so the difference is nonnegative, not negative. The claimed negative gap is therefore not established, and Theorem 1 as stated is false. The derivation also refers to 'x l, x\l are independent' without defining these symbols. Because Theorems 1 and 2 are used in Section III.C to motivate the learning objective, this is a load-bearing error; the theorem should be corrected (or removed) and the narrative adjusted so that the framework does not rest on an invalid bound.
  2. [Section IV.A.2 and Table VI] The block-based temporal split assigns the first and last 10% of each block to the test set and the middle 80% to training, but all blocks come from the same four VPFD participants. Training and test windows therefore share subject identity, sensor calibration, and activity context. With n=4 and a single fixed seed (Section IV.A.3), the reported improvements (e.g., 92.86% vs 82.31% for the Transformer) could reflect subject-specific or session-specific confounds rather than transferable fatigue-relevant knowledge. The central claim of cross-domain transfer requires leave-one-subject-out evaluation, per-subject results, or otherwise demonstrable generalization beyond the training subjects; the current protocol does not support the strong framing in the abstract and conclusion that the results demonstrate feasibility of transferring fatigue-relevant information across differently instrumented domains.
  3. [Appendix, Theorem 2 (Eq. 22)] The bound d_H∆H(P+, Q) <= d_H∆H(P, P) + d_H∆H(Qa, Qhat) <= d_H∆H(Qa, Qhat) is not justified. P+ and Q are joint distributions over (x,a) and (x,ahat), and their H∆H distance depends on the conditional distributions of a and ahat given x, not only on the marginal distributions Qa and Qhat. The triangle-inequality step with a zero 'P,P' term does not bridge these joint distributions unless additional structure such as a coupling or a conditional-distance bound is assumed. This weakens Theorem 2, which is the main theoretical justification for training the imputer by minimizing the regression loss between real and imputed modalities.
minor comments (5)
  1. [Section II.B] There is a typo: 'Early developments in the are primarily leveraged statistical regression techniques' should read 'Early developments primarily leveraged statistical regression techniques.'
  2. [Appendix, Definition 2 (Eq. 16)] The definition of the ideal classifier writes f* = arg min_f (E_P(f) + E_P(f)), but the two expectations should be over the two different distributions P and Q; otherwise the term E_P(f) is repeated.
  3. [Section IV.A.3] All experiments use a single fixed seed (42) and no error bars or significance tests are reported; given the small target cohort, reporting standard deviations over multiple seeds or bootstrap intervals would materially strengthen the empirical claims.
  4. [Table V] The row VPFD_MLP+noise reports a cross-entropy loss of 1.2041, which is lower than the VPFD baseline's 1.2789, yet its accuracy is much lower (60.58% vs 81.15%). This is not impossible with class imbalance or miscalibration, but it undercuts the stated interpretation that noise injection consistently degrades classifier performance; please clarify which metric supports the claim.
  5. [Author line and references] The author line contains a formatting artifact ('Tang Ying,Fellow, IEEE') and the FatigueSet dataset is cited as both [56] and [68]; please unify the references and correct the formatting.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: the empirical transfer claim is evaluated on held-out target-domain data using source-trained imputers, and theoretical bounds are motivational rather than fitted; the VPFD self-citation is a dataset reference, not load-bearing.

full rationale

The central claim—cross-domain imputation improves target fatigue detection—rests on external public datasets (MEFAR and FatigueSet) as sources and VPFD as target. Algorithm 2 trains the imputer g_s on source-domain shared modalities without target labels; Algorithm 1 then constructs enhanced target features and trains the classifier on target labels, with evaluation on temporally separated holdout windows from VPFD. No fitted parameter is renamed as a prediction, and the reported accuracy gain (e.g., 92.86% vs. 82.31% for the Transformer) is an empirical result over target holdout data. Theorems 1–3 are used only as motivation and are not fitted to data, so their mathematical flaws are correctness issues rather than circularity. The VPFD citation [46] is a self-citation, but it is a data reference to a public benchmark with its own collection protocol, not a load-bearing result that the present paper's conclusion reduces to. The acknowledged limitation of a small cohort (n=4) and the block-based temporal split raise generalization-risk concerns, but the paper does not define its outcome in terms of those choices. No step in the derivation chain is equivalent to its own input by construction.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The paper does not introduce new physical or model entities; imputed EEG/ECG are predictions of existing modalities, not new postulated constructs.

free parameters (4)
  • JacobianNorm regularization weight = 0.008
    Chosen by hand and used in all classification experiments; no sensitivity analysis is reported, and the ablation shows it changes accuracy by several points.
  • Resampling rate = 32 Hz
    All signals resampled to 32 Hz; this choice is arbitrary and affects all subsequent features.
  • Window length and sliding step = 10 s windows, 3 s step for training; 10 s non-overlapping for test
    Window size is manually selected; no ablation over window length is provided.
  • Noise baseline standard deviation = 2 times the max absolute gaze signal
    Ad hoc choice for the noise baseline in Section IV.B.2; it makes the baseline extremely noisy, which may exaggerate the value of imputation.
assumptions (5)
  • standard math R-sub-Gaussian loss and mutual information bound (Lemma III.1 from [63])
    Imported from Bu et al. 2020; used as the starting inequality for Theorems 1 and 2.
  • domain assumption Conditional independence of x and a given y (Theorem 1 assumption)
    Assumed without empirical evidence; required for the claimed bound but not verified on physiological data.
  • domain assumption Ideal classifier has small risk (Appendix, Assumption before Definition 3)
    Standard domain-adaptation assumption invoked for Theorem 2; not validated here.
  • ad hoc to paper BatchNorm reduces the HDeltaH distance between shared-feature distributions
    Used as the domain-alignment mechanism in the learning objective (Eq. 5); no formal or empirical support is given for this specific effect.
  • domain assumption Common physiological categories across different devices reflect the same underlying fatigue state
    Core transferability premise: HR from Empatica E4 and Pixel Watch 2 are treated as the same modality after scaling; the entire cross-domain imputation relies on this.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Enhancing Fatigue Detection through Heterogeneous Multi-Source Data Integration and Cross-Domain Modality Imputation." pith.science (2026). https://pith.science/paper/4QR5WXNY

@misc{pith2026250716859,
  author       = {Pith},
  title        = {Pith review of: Enhancing Fatigue Detection through Heterogeneous Multi-Source Data Integration and Cross-Domain Modality Imputation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4QR5WXNY}},
  note         = {Machine review of arXiv:2507.16859}
}
read the original abstract

Fatigue detection for human operators is important in safety-related applications such as aviation, mining, and long-haul transport. Reliable estimation of operator fatigue can support timely warnings, adaptive task scheduling, takeover reminders, and other safety-management decisions in human-machine systems. However, the effectiveness of these functions depends on whether fatigue-related signals can be reliably captured in the deployment environment. While many studies have shown the value of high-fidelity sensors in controlled laboratory environments, their performance often degrades when used in real-world settings because of noise, lighting conditions, and field-of-view constraints, thereby limiting their practical use. This paper formalizes a deployment-oriented setting for real-world fatigue detection, where high-quality sensors are often unavailable in practical applications. To address this issue, we use knowledge from heterogeneous source domains, including high-fidelity sensors that are difficult to deploy in the field but commonly used in controlled environments, to assist fatigue detection in the real-world target domain. Based on this idea, we design a heterogeneous and multi-source fatigue-detection framework that uses the available modalities in the target domain while leveraging diverse configurations in the source domains through cross-domain modality imputation based on shared modalities.

Figures

Figures reproduced from arXiv: 2507.16859 by the authors.

Figure 1
Figure 1. High-level overview of the proposed framework. Blue solid boxes denote available modalities observed in source or target domains, and orange [PITH_FULL_IMAGE:figures/full_fig_p007_1.png] view at source ↗
Figure 3
Figure 3. Modality imputation module. Shared modalities such as HR, PPG, [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Fatigue detection model. For VPFD, the classifier uses both available [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

75 extracted references · 55 canonical work pages

  1. [1]

    Hooda, V

    R. Hooda, V . Joshi, M. Shah, A comprehensive review of ap- proaches to detect fatigue using machine learning techniques, Chronic Diseases and Translational Medicine 8 (1) (2022) 26–35. doi:10.1016/j.cdtm.2021.07.002

  2. [2]

    Zhang, Q

    Y . Zhang, Q. Gao, R. Hu, Q. Ding, B. Li, Y . Guo, Differentiable prior- driven data augmentation for sensor-based human activity recognition, IEEE Transactions on Computational Social Systems 12 (5) (2025) 3778–3790. doi:10.1109/TCSS.2025.3565414

  3. [3]

    J. J. Bird, M. Pritchard, A. Fratini, A. Ek ´art, D. R. Faria, Synthetic bio- logical signals machine-generated by gpt-2 improve the classification of eeg and emg through data augmentation, IEEE Robotics and Automation Letters 6 (2) (2021) 3498–3504. doi:10.1109/LRA.2021.3056355

  4. [4]

    A. G. Habashi, A. M. Azab, S. Eldawlatly, G. M. Aly, Motor imagery classification enhancement using generative adversarial networks for eeg spectrum image generation, in: 2023 IEEE 36th International Sympo- sium on Computer-Based Medical Systems (CBMS), 2023, pp. 354–359. doi:10.1109/CBMS58004.2023.00243

  5. [5]

    H. Zeng, X. Li, G. Borghini, Y . Zhao, P. Aric `o, G. Di Flumeri, N. Scia- raffa, W. Zakaria, W. Kong, F. Babiloni, An eeg-based transfer learning method for cross-subject fatigue mental state prediction, Sensors 21 (7) (2021) 2369

  6. [6]

    Zhang, M

    X. Zhang, M. Zeman, T. Tsiligkaridis, M. Zitnik, Graph-guided network for irregularly sampled multivariate time series, International Conference on Learning Representations, ICLR (2022)

  7. [7]

    J. Hu, P. Wang, Noise robustness analysis of performance for eeg-based driver fatigue detection using different entropy feature sets, Entropy 19 (8) (2017). doi:10.3390/e19080385. URL https://www.mdpi.com/1099-4300/19/8/385

  8. [8]

    Borghini, G

    G. Borghini, G. Vecchiato, J. Toppi, L. Astolfi, A. Maglione, R. Isabella, C. Caltagirone, W. Kong, D. Wei, Z. Zhou, L. Polidori, S. Vitiello, F. Ba- biloni, Assessment of mental fatigue during car driving by using high resolution eeg activity and neurophysiologic indices, in: 2012 Annual International Conference of the IEEE Engineering in Medicine and Bi...

Show all 75 references
  1. [9]

    Sikander, S

    G. Sikander, S. Anwar, Driver fatigue detection systems: A review, IEEE Transactions on Intelligent Transportation Systems 20 (6) (2019) 2339–

  2. [10]

    Jim ´enez-Guarneros, G

    M. Jim ´enez-Guarneros, G. Fuentes-Pineda, J. Grande-Barreto, Mmda: A multimodal and multisource domain adaptation method for cross- subject emotion recognition from eeg and eye movement signals, IEEE Transactions on Computational Social Systems 12 (5) (2025) 2214–

  3. [11]

    Shahid, K

    A. Shahid, K. Wilkinson, S. Marcu, C. M. Shapiro, Stanford sleepiness scale (sss), in: A. Shahid, K. Wilkinson, S. Marcu, C. M. Shapiro (Eds.), STOP, THAT and One Hundred Other Sleep Scales, Springer New York, New York, NY , 2012, pp. 369–370

  4. [12]

    Smets, B

    E. Smets, B. Garssen, B. Bonke, J. De Haes, The multidimensional fatigue inventory (mfi) psychometric qualities of an instrument to assess fatigue, Journal of psychosomatic research 39 (3) (1995) 315–25

  5. [13]

    F. B. Ortega, J. R. Ruiz, V . Espa ˜na-Romero, G. Vicente-Rodriguez, D. Mart ´ınez-G´omez, Y . Manios, L. B ´eghin, D. Molnar, K. Widhalm, L. A. Moreno, M. Sj¨ostr¨om, M. J. Castillo, The international fitness scale (ifis): usefulness of self-reported fitness in youth, Interna...

  6. [14]

    V . J. Gawron, Overview of self-reported measures of fatigue, The International Journal of Aviation Psychology 26 (3-4) (2016) 120–131

  7. [15]

    Brown, S

    N. Brown, S. Bichler, M. Fiedler, W. Alt, Fatigue detection in strength training using three-dimensional accelerometry and principal component analysis, Sports biomechanics 15 (2) (2016) 139–150

  8. [16]

    B. T. Jap, S. Lal, P. Fischer, E. Bekiaris, Using eeg spectral compo- nents to assess algorithms for detecting fatigue, Expert Systems with Applications 36 (2) (2009) 2352–2359

  9. [17]

    S. Kar, M. Bhagat, A. Routray, Eeg signal analysis for the assessment and quantification of driver’s fatigue, Transportation research part F: traffic psychology and behaviour 13 (5) (2010) 297–306

  10. [18]

    S. Fu, Z. Yang, Y . Ma, Z. Li, L. Xu, H. Zhou, Advancements in the intelligent detection of driver fatigue and distraction: A comprehensive review, Applied Sciences 14 (7) (2024). doi:10.3390/app14073016

  11. [19]

    Olaganathan, T

    R. Olaganathan, T. Holt, J. Luedtke, B. Bowen, Fatigue and its manage- ment in the aviation industry, with special reference to pilots, Journal of Aviation Technology and Engineering 10 (2021) 45. doi:10.7771/2159- 6670.1208

  12. [20]

    E. Q. Wu, Y . Gao, W. Tong, Y . Hou, R. Law, G. Zhu, Cognitive state detection in task context based on graph attention network during flight, IEEE Transactions on Systems, Man, and Cybernetics: Systems 54 (9) (2024) 5224–5236. doi:10.1109/TSMC.2024.3380078

  13. [21]

    G. Li, W. Yan, S. Li, X. Qu, W. Chu, D. Cao, A temporal–spatial deep learning approach for driver distraction detection based on eeg signals, IEEE Transactions on Automation Science and Engineering 19 (4) (2022) 2665–2677. doi:10.1109/TASE.2021.3088897

  14. [22]

    Y . Yang, Z. Gao, Y . Li, Q. Cai, N. Marwan, J. Kurths, A complex network based broad learning system for detecting driver fatigue from eeg signals, IEEE Transactions on Systems, Man, and Cybernetics: Systems 51 (9) (2021) 5800–5808. doi:10.1109/TSMC.2019.2956022

  15. [23]

    H. T. Nguyen, N. D. Mai, B. G. Lee, W.-Y . Chung, Behind-the-ear eeg- based wearable driver drowsiness detection system using embedded tiny neural networks, IEEE Sensors Journal 23 (19) (2023) 23875–23892. doi:10.1109/JSEN.2023.3307766

  16. [24]

    Chand, A

    S. Chand, A. J. McDaid, Y . Lu, Isometric-based approach for de- tecting localized muscular fatigue during complex dynamic manufac- turing operations, in: 2021 IEEE 17th International Conference on Automation Science and Engineering (CASE), 2021, pp. 1940–1945. doi:10.1109/CAS...

  17. [25]

    Al-Libawy, A

    H. Al-Libawy, A. Al-Ataby, W. Al-Nuaimy, M. A. Al-Taee, Hrv-based operator fatigue analysis and classification using wearable sensors, in: 2016 13th International Multi-Conference on Systems, Signals & Devices (SSD), 2016, pp. 268–273. doi:10.1109/SSD.2016.7473750

  18. [26]

    Markiewicz, A

    R. Markiewicz, A. Markiewicz-Gospodarek, B. Dobrowolska, Galvanic skin response features in psychiatry and mental disorders: A narrative review, International Journal of Environmental Research and Public Health 19 (20) (2022) 13428. doi:10.3390/ijerph192013428

  19. [27]

    Y . Jiao, C. Zhang, X. Chen, L. Fu, C. Jiang, C. Wen, Driver fatigue detection using measures of heart rate variability and electrodermal activity, IEEE Transactions on Intelligent Transportation Systems 25 (6) (2024) 5510–5524. doi:10.1109/TITS.2023.3333252

  20. [28]

    Mohanavelu, R

    K. Mohanavelu, R. Lamshe, S. Poonguzhali, K. Adalarasu, M. Jagan- nath, Assessment of human fatigue during physical performance using physiological signals: A review, Biomedical and Pharmacology Journal 10 (4) (2017) 1887–1896

  21. [29]

    B. Hu, L. Zhu, Q. Dong, K. Qian, X. Wang, J. Li, Q. Zhang, J. Yan, X. Wang, Y . Zhang, X. Zhang, Y . Qi, P. Li, W. Zheng, J. Li, J. Chen, 12 X. Li, C. Shi, W. Sun, S. Li, G. Zhao, H. Yin, L. Sun, H. Peng, X. Gao, D. Zhang, S. Gao, C.-T. Lin, L. Tong, F. Chen, J. Chen, C. Wang,...

  22. [30]

    Cheng, C

    B.-Q. Cheng, C. Fan, H. Fu, J. Huang, H. Chen, X. Luo, Mea- suring and computing cognitive statuses of construction workers based on electroencephalogram: A critical review, IEEE Transac- tions on Computational Social Systems 9 (6) (2022) 1644–1659. doi:10.1109/TCSS.2022.3158585

  23. [31]

    L. J. Trejo, K. Kubitz, R. Rosipal, R. L. Kochavi, L. D. Montgomery, Eeg-based estimation and classification of mental fatigue, Psychology 6 (5) (2015) 572–589

  24. [32]

    Cheng, H.-T

    S.-Y . Cheng, H.-T. Hsu, Mental fatigue measurement using eeg, In Risk Management Trends (2011)

  25. [33]

    Li, W.-Y

    G. Li, W.-Y . Chung, Smartwatch-based wearable eeg system for driver drowsiness detection, IEEE Sensors Journal 15 (12) (2015) 7169–7180. doi:10.1109/JSEN.2015.2473679

  26. [34]

    G. Chen, L. Hong, J. Dong, P. Liu, J. Conradt, A. Knoll, Eddd: Event- based drowsiness driving detection through facial motion analysis with neuromorphic vision sensor, IEEE Sensors Journal 20 (11) (2020) 6170–

  27. [35]

    LaRocco, M

    J. LaRocco, M. D. Le, D.-G. Paeng, A systemic review of available low-cost eeg headsets used for drowsiness detection, Frontiers in neu- roinformatics 14 (2020) 553352

  28. [36]

    Shashidhar, S

    R. Shashidhar, S. Hariprasad, B. SHRUTHI, R. S. KA VITA, Real-time fatigue detection using a low cost wireless eeg device, i-Manager’s Journal on Embedded Systems 8 (2) (2020) 8–13

  29. [37]

    F. Liu, D. Chen, J. Zhou, F. Xu, A review of driver fatigue detection and its advances on the use of rgb-d camera and deep learning, Engineering Applications of Artificial Intelligence 116 (2022) 105399. doi:https://doi.org/10.1016/j.engappai.2022.105399

  30. [38]

    Y . Yi, Z. Zhou, W. Zhang, M. Zhou, Y . Yuan, C. Li, Fatigue detection algorithm based on eye multifeature fusion, IEEE Sensors Journal 23 (7) (2023) 7949–7955. doi:10.1109/JSEN.2023.3247582

  31. [39]

    W. Deng, R. Wu, Real-time driver-drowsiness detection system using facial features, IEEE Access 7 (2019) 118727–118738. doi:10.1109/ACCESS.2019.2936663

  32. [40]

    Zheng, S

    H. Zheng, S. Chand, A. Keshvarparast, D. Battini, Y . Lu, Video-based fatigue estimation for human-robot task allocation optimisation, in: 2023 IEEE 19th International Conference on Automation Science and Engineering (CASE), 2023, pp. 1–6. doi:10.1109/CASE56687.2023.10260460

  33. [41]

    X. Li, W. Xu, B. Yao, Z. Ji, X. Liu, Dynamic task reallocation in human-robot collaborative workshop based on online biotic fa- tigue detection, in: 2022 IEEE 18th International Conference on Automation Science and Engineering (CASE), 2022, pp. 116–122. doi:10.1109/CASE49997.2...

  34. [42]

    Yanqing, D

    S. Yanqing, D. Pengfei, Z. Yangjing, C. Shitao, Z. Nanning, Privacy protection of sensitive bioinformation based on event cameras, Strategic Study of CAE 26 (1) (2024). doi:10.15302/J-SSCAE-2024.01.017

  35. [43]

    Krejtz, A

    K. Krejtz, A. T. Duchowski, A. Niedzielska, C. Biele, I. Krejtz, Eye tracking cognitive load using pupil diameter and microsaccades with fixed gaze, PloS one 13 (9) (2018) e0203629

  36. [44]

    F. N. Biondi, F. Graf, J. Cort, Testing pupil size as a possible alternative metric of physical fatigue in automotive manufacturing tasks, Proceed- ings of the Human Factors and Ergonomics Society Annual Meeting 67 (1) (2023) 824–828. doi:10.1177/21695067231192895

  37. [45]

    N. Y . Yair Morad, Hadas Lemberg, Y . Dagan, Pupillography as an objective indicator of fatigue, Current Eye Research 21 (1) (2000) 535–

  38. [46]

    L. Cui, Y . Wu, Y . Tang, Generalizable gaze synthesis for prac- tical fatigue detection systems, in: 2024 International Conference on Cyber-Physical Social Intelligence (ICCSI), 2024, pp. 1–6. doi:10.1109/ICCSI62669.2024.10799435

  39. [47]

    Neifar, A

    N. Neifar, A. Mdhaffar, A. Ben-Hamadou, M. Jmaiel, Deep generative models for physiological signals: A systematic literature review, Artificial Intelligence in Medicine 165 (2025) 103127. doi:https://doi.org/10.1016/j.artmed.2025.103127. URL https://www.sciencedirect.com/scien...

  40. [48]

    Vetter, J

    J. Vetter, J. H. Macke, R. Gao, Generating realistic neurophysiological time series with denoising diffusion probabilistic models, Patterns 5 (8) (2024) 101047. doi:10.1016/j.patter.2024.101047

  41. [49]

    Z. Che, S. Purushotham, K. Cho, D. Sontag, Y . Liu, Recurrent neural networks for multivariate time series with missing values, in: Proceed- ings of the 2016 ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, ACM, 2016, pp. 749–758

  42. [50]

    Verma, A

    A. Verma, A. Kumar, An accurate missing data prediction method using lstm based deep learning for health care, Procedia Computer Science 167 (2019) 2414–2423

  43. [51]

    W. Cao, D. Wang, J. Li, H. Zhou, S. Li, Y . He, Y . Liu, H. Zha, Brits: Bidirectional recurrent imputation for time series, arXiv preprint arXiv:1805.10572 (2018)

  44. [52]

    J. Shen, K. Zhu, H. Liu, J. Wu, K. Wang, Q. Dong, Tensor correlation fusion for multimodal physiological signal emotion recognition, IEEE Transactions on Computational Social Systems 11 (6) (2024) 7299–

  45. [53]

    Eum, E.-H

    Y . Eum, E.-H. Yoo, Imputation of missing time-activity data with long-term gaps: A multi-scale residual cnn-lstm network model, Computers, Environment and Urban Systems 95 (2022) 101823. doi:10.1016/j.compenvurbsys.2022.101823

  46. [54]

    W. Li, C. Gu, J. Chen, C. Ma, X. Zhang, B. Chen, S. Wan, Dls-gan: Generative adversarial nets for defect location sensitive data augmenta- tion, IEEE Transactions on Automation Science and Engineering 21 (4) (2024) 5173–5189. doi:10.1109/TASE.2023.3309629

  47. [55]

    A. F. Nia, V . Tang, G. M. Talou, M. Billinghurst, Synthesizing affective neurophysiological signals using generative models: A review paper, Journal of Neuroscience Methods 406 (2024) 110129. doi:https://doi.org/10.1016/j.jneumeth.2024.110129. URL https://www.sciencedirect.co...

  48. [56]

    Kalanadhabhatta, C

    M. Kalanadhabhatta, C. Min, A. Montanari, F. Kawsar, Fatigueset: A multi-modal dataset for modeling mental fatigue and fatigability, in: H. Lewy, R. Barkan (Eds.), Pervasive Computing Technologies for Healthcare, Springer International Publishing, Cham, 2022, pp. 204– 217

  49. [57]

    Derdiyok, F

    S. Derdiyok, F. P. Akbulut, C. Catal, Neurophysiological and biosignal data for investigating occupational mental fatigue: Mefar dataset, Data in Brief 52 (2024) 109896. doi:https://doi.org/10.1016/j.dib.2023.109896

  50. [58]

    Papakostas, A

    M. Papakostas, A. Rajavenkatanarayanan, F. Makedon, Cogbeacon: A multi-modal dataset and data-collection platform for modeling cognitive fatigue, Technologies 7 (2) (2019). doi:10.3390/technologies7020046

  51. [59]

    Markova, T

    V . Markova, T. Ganchev, K. Kalinkov, Clas: A database for cognitive load, affect and stress recognition, in: 2019 International Conference on Biomedical Innovations and Applications (BIA), 2019, pp. 1–4. doi:10.1109/BIA48344.2019.8967457

  52. [60]

    Gabbi, L

    M. Gabbi, L. Cornia, V . Villani, L. Sabattini, Understanding fatigue through biosignals: A comprehensive dataset, 2024 19th ACM/IEEE International Conference on Human-Robot Interaction (HRI) (2024) 901–905

  53. [61]

    Kovalenko, A

    S. Kovalenko, A. Mamonov, V . Kuznetsov, A. Bulygin, I. Shoshina, I. Brak, A. Kashevnik, Operatoreyevp: Operator dataset for fatigue de- tection based on eye movements, heart rate data, and video information, Sensors 23 (13) (2023). doi:10.3390/s23136197

  54. [62]

    Schmidt, A

    P. Schmidt, A. Reiss, R. Duerichen, C. Marberger, K. Van Laer- hoven, Introducing wesad, a multimodal dataset for wearable stress and affect detection, in: Proceedings of the 20th ACM International Conference on Multimodal Interaction, ICMI ’18, Association for Computing Machi...

  55. [63]

    Y . Bu, S. Zou, V . V . Veeravalli, Tightening Mutual Informa- tion Based Bounds on Generalization Error, IEEE Journal on Selected Areas in Information Theory 1 (1) (2020) 121–130. doi:10.1109/JSAIT.2020.2991139

  56. [64]

    Mansour, M

    Y . Mansour, M. Mohri, A. Rostamizadeh, Domain adaptation: Learning bounds and algorithms, arXiv preprint arXiv:0902.3430 (2009)

  57. [65]

    Ioffe, C

    S. Ioffe, C. Szegedy, Batch normalization: Accelerating deep network training by reducing internal covariate shift (2015). arXiv:1502.03167. URL https://arxiv.org/abs/1502.03167

  58. [66]

    Y . Li, N. Wang, J. Shi, J. Liu, X. Hou, Revisiting batch normalization for practical domain adaptation (2016). arXiv:1603.04779. URL https://arxiv.org/abs/1603.04779

  59. [67]

    W. Li, S. Chen, Partial domain adaptation without domain alignment, IEEE Transactions on Pattern Analysis and Machine Intelligence 45 (7) (2023) 8787–8797. doi:10.1109/tpami.2022.3228937. URL http://dx.doi.org/10.1109/TPAMI.2022.3228937

  60. [68]

    Kalanadhabhatta, C

    M. Kalanadhabhatta, C. Min, A. Montanari, F. Kawsar, Fatigueset: A multi-modal dataset for modeling mental fatigue and fatigability, in: International Conference on Pervasive Computing Technologies for Healthcare, Springer, 2021, pp. 204–217

  61. [69]

    S. Wang, H. Wu, X. Shi, T. Hu, H. Luo, L. Ma, J. Y . Zhang, J. Zhou, Timemixer: Decomposable multiscale mixing for time series forecasting 13 (2025). arXiv:2405.14616. URL https://arxiv.org/abs/2405.14616

  62. [70]

    Vaswani, N

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser, I. Polosukhin, Attention is all you need, in: Advances in Neural Information Processing Systems, V ol. 30, 2017, pp. 5998–6008. LuoBin Cui(Member, IEEE) received the B.S. degree from Jiangsu Uni...

  63. [542]

    doi:10.1076/0271-3683(200007)2111-ZFT535

  64. [2227]

    doi:10.1109/TCSS.2024.3519300

  65. [2352]

    doi:10.1109/TITS.2018.2868499

  66. [6180]

    doi:10.1109/JSEN.2020.2973049

  67. [7308]

    doi:10.1109/TCSS.2024.3406988

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.